A Markov Decision Process Model for a Reinforcement Learning-based Autonomous Pedestrian Crossing Protocol

Autonomous Traffic Management (ATM) systems empowered with Machine Learning (ML) technics are a promising solution for eliminating traffic light and decreasing traffic congestion in the future. However, few efforts have focused on integrating pedestrians in ATM, namely the static programming-based cooperative protocol called Autonomous Pedestrian Crossing (APC). In this paper, we model a Markov Decision Process (MDP) to enable a Deep Reinforcement Learning (DRL)-based version of APC protocol that is able to dynamically achieve the same objectives (i.e. decreasing traffic delay at the crossing area). Using concrete state space, action set and reward functions, our model forces the Autonomous Vehicle (AV) to "think" and behave according to APC architecture. Compared to the traditional programming APC system, our approach permits the AV to learn from its previous experiences in non-signalized crossing and optimize the distance and the velocity parameters accordingly.